What's new

Gemini 3.7 Flash completed all 35 builds with a 3.00 average across graded samples, showing solid native canvas physics alongside external library loading issues.

CLEAN

gemini-3.7-flash

CLEAN

gemini-3.7-flash

CLEAN

gemini-3.7-flash

CLEAN

gemini-3.7-flash

We added the smaller Qwen3.8-27B, so we put it straight up against the 2.4-trillion-parameter Qwen3.8-2.4T-A95B already in the gallery — same family, roughly two orders of magnitude apart in size. Both are reliable builders (34 of 35 and 35 of 35 shipped), so the story is entirely in the quality.

BLANK

qwen3.8-27b

CLEAN

qwen3.8-2.4t-a95b

BLANK

qwen3.8-27b

CLEAN

qwen3.8-2.4t-a95b

xAI's Grok 4.6 ran the full matrix clean — 35 of 35 builds shipped — and averaged 3.43, with a clean five and a clear strength in generative canvas work.

CLEAN

grok-4.6

CLEAN

grok-4.6

CLEAN

grok-4.6

CLEAN

grok-4.6

Meta shipped two new Muse models, so we ran both through the same 35 one-shot prompts. Neither failed to ship a single build — 35 of 35 runnable each — so the gap between them is entirely about quality, not reliability.

CLEAN

muse-spark-1.2

CLEAN

muse-glimmer-30b

CLEAN

muse-spark-1.2

CLEAN

muse-glimmer-30b

Alibaba's 2.4-trillion-parameter Qwen3.8-2.4T-A95B turned in the most consistent run of the batch — 35 of 35 builds shipped, zero blank or crashed, and the tightest quality spread in the set.

CLEAN

qwen3.8-2.4t-a95b

CLEAN

qwen3.8-2.4t-a95b

CLEAN

qwen3.8-2.4t-a95b

CLEAN

qwen3.8-2.4t-a95b

Two new ByteDance Seed models went through the matrix together — both 35 of 35 on build reliability, but with opposite temperaments.

CLEAN

seed-2.0-code

CLEAN

seed-2-1-turbo

CLEAN

seed-2.0-code

CLEAN

seed-2-1-turbo

Alibaba's new flagship Qwen3.8-Max ran all 35 one-shot prompts with zero build failures — a clean sweep of the matrix, and a solid mid-pack showing on quality.

CLEAN

qwen3.8-max

CLEAN

qwen3.8-max

CLEAN

qwen3.8-max

CLEAN

qwen3.8-max

A new side-by-side grid: pick up to 10 models and see how each one built every prompt, in one scrollable table.

We ran Moonshot's new Kimi K3 and Claude Opus 4.8 through the same 34 one-shot prompts, same harness, and they finish within a hair of each other — K3 at 3.24 average, Opus at 3.12. The interesting part is where each one misses.

CLEAN

kimi-k3

PARTIAL

claude-opus-4.8

CLEAN

kimi-k3

BLANK

claude-opus-4.8

Moonshot shipped Kimi K3, a million-token model, so we added it to the matrix and ran all 34 one-shot prompts against it — same harness, same single-shot rules as every other build in the gallery.

CLEAN

kimi-k3

CLEAN

kimi-k3

CLEAN

kimi-k3

CLEAN

kimi-k3

Five more models joined the matrix this week, and we re-ran every prompt against them. The alien-shooter build is a good stress test: it needs a game loop, collision, and input handling all in one file, so the gap between models shows up fast.

CLEAN

deepseek-v4-pro

PARTIAL

kimi-k2.7-code

BROKEN

glm-5.2

BLANK

minimax-m2.7

Every build now carries a two-tier grade. Tier 1 measures the artifact deterministically — how much of the screen moves, whether it responds to input, console and JS errors — and Tier 2 hands those measurements plus a capture to a vision model for the final verdict.

CLEAN

claude-opus-4.8

CLEAN

deepseek-v4-pro

PARTIAL

claude-opus-4.8

CLEAN

deepseek-v4-flash

A real-time WebGL fluid sim in one file — the hardest prompt in the set so far.

CLEAN

deepseek-v4-pro

CLEAN

glm-5

CLEAN

minimax-m2.7

PARTIAL

claude-opus-4.8